TF_NewWritableFile(full_path, &h, status);
  if (TF_GetCode(status) != TF_OK) {
    fprintf(stderr, "TF_NewWritableFile failed: %s\n", TF_Message(status));
    return 1;
  }
  fprintf(stderr, "wrote %s\n", full_path);
  free(full_path);
  TF_CloseWritableFile(h, status);
  if (TF_GetCode(status) != TF_OK) {
    fprintf(stderr, "TF_CloseWritableFile failed: %s\n", TF_Message(status));
  }

      // Release all backprop functions
      for (const auto& pair : state.op_tape) {
        pair.second.backward_function_deleter(pair.second.backward_function);
      }
    }
  });
  if (!s.ok()) {
    return s;
  }

  std::unordered_map<int64_t, int64_t> gradients_size;
  // TODO(apassos) multiple threads could be dequeuing from op_stack at the same
  // time, for better CPU backprop performance.